Computed tomography using meta-optics

TL;DR

A metaoptic imager is presented, which implements the Radon transform obviating the need for training the optics and high quality image reconstruction with a large compression ratio is presented through the use of the Simultaneous Algebraic Reconstruction Technique.

Abstract

Computer vision tasks require processing large amounts of data to perform image classification, segmentation, and feature extraction. Optical preprocessors can potentially reduce the number of floating point operations required by computer vision tasks, enabling low-power and low-latency operation. However, existing optical preprocessors are mostly learned and hence strongly depend on the training data, and thus lack universal applicability. In this paper, we present a metaoptic imager, which implements the Radon transform obviating the need for training the optics. High quality image reconstruction with a large compression ratio of 0.6% is presented through the use of the Simultaneous Algebraic Reconstruction Technique. Image classification with 90% accuracy is presented on an experimentally measured Radon dataset through neural network trained on digitally transformed images.

Figures (5)

• Figure 1: a. Schematic for meta-optic computed tomography. An object is imaged with cylindrical metalens at the back-focal-plane along the focal line, measuring the DC intensity component of 1-D Fourier transforms along different angles corresponding to different slices in Radon space. b. Schematic of Radon transform operation. Integrals along straight lines $L \subset \mathbb{R}^2$ are taken at different rotation angles $\theta$ to construct $\mathcal{R}f$c. Comparison of cost vs number of pixels of different imaging systems. Our system is low cost and can achieve imaging with a very high number of pixels.
• Figure 2: A. Scanning electron micrographs (SEM) of fabricated metalens. Scale bars correspond to $5\mu m$ for the left most SEM image, $5\mu m$ for the center, and $1\mu m$ for the rightmost. B. Experimental setup. A display projects an image through a pinhole. A lens is used to collimate light from the pinhole onto the cylindrical metasurface. The Cylindrical metasurface and the detector are mounted onto a rotating stage, which rotates the system to measure the image at different angles.
• Figure 3: Experimental results of metalens-based imaging using the radon transform. The left column is the true scene being imaged. Column second from the left are the raw measured sinograms collected experimentally. The x axis are the coordinates of the projection angle, and the y axis are the measured pixel values. The four rightmost columns are different steps in the reconstruction algorithm. Algorithm converges at about 100 iterations.
• Figure 4: A. The setup of the two imaging systems. Top shows a regular scene and detector system and an ideal output. Bottom shows the proposed Radon imaging system. B. Neural network architectures. Top conventional AlexNet based convolutions neural network with 3 convolutional layers, and 2 densely connected layers. Bottom Radon transform based neural network. The proposed layers are thinner due to the measured Radon transform containing fewer pixels in the angular dimension.
• Figure 5: Confusion matrices of the (left) validation set with radon transform computed numerically, and (right) experimentally measured radon transform.